Predicting Regioselectivity and Lability of Cytochrome P450

Article pubs.acs.org/jcim

Predicting Regioselectivity and Lability of Cytochrome P450 Metabolism Using Quantum Mechanical Simulations Jonathan D. Tyzack, Peter A. Hunt, and Matthew D. Segall* Optibrium Ltd., 7221 Cambridge Research Park, Beach Drive, Cambridge CB25 9TL, U.K. S Supporting Information *

ABSTRACT: We describe methods for predicting cytochrome P450 (CYP) metabolism incorporating both pathway-specific reactivity and isoform-specific accessibility considerations. Semiempirical quantum mechanical (QM) simulations, parametrized using experimental data and ab initio calculations, estimate the reactivity of each potential site of metabolism (SOM) in the context of the whole molecule. Ligand-based models, trained using high-quality regioselectivity data, correct for orientation and steric effects of the different CYP isoform binding pockets. The resulting models identify a SOM in the top 2 predictions for between 82% and 91% of compounds in independent test sets across seven CYP isoforms. In addition to predicting the relative proportion of metabolite formation at each site, these methods estimate the activation energy at each site, from which additional information can be derived regarding their lability in absolute terms. We illustrate how this can guide the design of compounds to overcome issues with rapid CYP metabolism.

■

INTRODUCTION

The catalytic action of CYPs is predominantly that of a monooxygenase:

The development of methods to predict the sites, products, and rates of metabolism is an important avenue of research and finds application in the development of drugs, cosmetics, nutritional supplements, and agrochemicals. It is necessary to understand the pharmacokinetics of a molecule and ensure that it has sufficient exposure at the target to exert its therapeutic effect. In this regard it would be helpful to give an absolute prediction of the rate of, or at least the lability of a compound to, metabolism, rather than just a rank ordering of sites; a factor often neglected by metabolism prediction tools. It is also important to predict the formation of toxic metabolites, which contributes to the high attrition rates experienced in the development of new chemical entities, the imposition of black-box warnings, or even the withdrawal of approved pharmaceuticals. Thus, the ability to identify potential toxic metabolites early and make predictions about metabolic stability is of crucial importance in the drug discovery process. The cytochrome P450s (CYPs) are a family of hemecontaining enzymes involved in the phase-I metabolism of over 90% of drugs currently on the market.1 The CYP family consists of 57 isoforms2 with the largest contribution to xenobiotic metabolism coming from CYP3A4, the most promiscuous isoform, followed by CYP2D6 and CYP2C9. A comprehensive overview of the structure, reactivity, and catalytic cycle of CYPs can be found in the review paper from Shaik et al.3 © XXXX American Chemical Society

RH + O2 + 2H+ + 2e− → ROH + H 2O

(1)

where RH is the substrate molecule. The most common reactions catalyzed by CYPs involve the insertion of a single oxygen into an organic molecule, such as CC epoxidation, aromatic C-oxidation, and aliphatic C-hydroxylation, the last example often leading to N-dealkylation or O-dealkylation if oxidation occurs on a suitable leaving group in an amine or ether moiety. The addition of oxygen into the substrate is a precursor to excretion from the body, driving an increase in polarity and hydrophilicity and facilitating Phase II metabolism pathways such as glucuronidation. The heme moiety at the catalytic center of the CYPs is conserved across all isoforms, where a highly activated oxy-heme, formed by cleavage of molecular oxygen, is generated within the catalytic cycle, as shown in Figure 1. In addition to the main catalytic cycle there are two significant decoupling pathways, labeled D1 and D2, resulting in the formation of hydrogen peroxide and water, respectively, and returning the active-site heme to an inactivated state. The relative rate of decoupling compared to substrate metabolism will influence the observed rate of the conversion of substrate into metabolites and is an Received: April 26, 2016 Published: October 18, 2016 A

DOI: 10.1021/acs.jcim.6b00233 J. Chem. Inf. Model. XXXX, XXX, XXX−XXX

Article

Journal of Chemical Information and Modeling

each as a discrete uniform entity regardless of neighboring functional groups • QM methods can estimate the activation energy of the rate-limiting step of the oxidation reaction, allowing comparison of lability on an absolute scale. The previously published SMARTCyp method15 also uses a QM-based approach to predict CYP SOM. However, the approach differs from that described herein in that the SOMs are ranked based on a look-up table of small functionalities for which the activation energies have been previously calculated using ab initio density functional (DFT) methods. The use of ab initio QM methods avoids the need for detailed experimental data on which to base estimates of the activation energies. However, these calculations are computationally very expensive, and the use of a look-up of pre-calculated results is required to return results in a practical time frame. Therefore, this approach does not take into account potential long-range effects due to the environment of the whole molecule, an important factor for a medicinal chemist developing a lead series and aiming to predict the likely impact of structural changes on metabolic stability. The use of semiempirical QM calculations that estimate the activation energies for each aliphatic and aromatic SOM has previously been described.16 Semiempirical methods are significantly faster than ab initio methods and therefore can be applied to an entire substrate on a routine basis. However, they typically require detailed experimental data with which to parametrize a free energy relationship, and therefore these models do not include less common, but important, pathways such as epoxidation or N- and S-oxidation, due to the lack of sufficient experimental data. The methods described herein build on both of these methods to achieve transferability, application to the whole substrate to explicitly consider the molecular environment of each SOM and computational efficiency, returning results in approximately 1−2 min per compound on a single CPU. In the following section, the theory and implementation of the models will be explained in detail. The performance of the models on independent test sets will be presented in the Results section with comparisons made to the SMARTCyp15 method. Finally, three example applications will be presented to illustrate how the models can provide valuable guidance to redesign compounds and overcome issues due to rapid CYP metabolism.

Figure 1. Catalytic cycle of the cytochrome P450 enzymes. The decoupling pathways to form hydrogen peroxide and water are labeled D1 and D2, respectively.

important consideration when making absolute assessments of metabolic stability. Experimental investigation of xenobiotic metabolism can be both resource and time consuming, which has encouraged the development of computational techniques. These can be separated into two distinct categories: ligand-based and structure-based. In the first approach, structures and properties of known substrate or non-substrate ligands are modeled to develop structure−activity relationships. The second approach is focused on the structure of the metabolizing CYP enzyme, its known reaction mechanisms, and its interactions with substrates. Structure-based methods include docking, to identify potential ligand conformations in the context of the enzyme active site;4 molecular dynamics simulations, which can estimate the energetics of binding;5,6 and QM/MM methods, which combine a quantum mechanical description of the catalytic reaction center while capturing the effects of the surrounding protein structure using an empirical molecular mechanics force field.7 Structurebased methods can provide useful insight into the threedimensional influence of the protein structure on the site of reaction and binding affinity, but they are limited by the availability of good protein structures from crystallography and the flexibility of many P450 isoforms, in particular CYP3A4, which is problematic to take into consideration in a reasonable time frame. The available evidence also suggests that structurebased methods do not, at present, improve the accuracy of predictions over ligand-based approaches.8,9 For a general overview of current computational tools to predict sites of metabolism (SOM), the reader is referred to the many comprehensive review papers.10−14 Most metabolism prediction tools incorporate some form of reactivity and accessibility considerations. The method described herein is no exception: a ligand-based approach is used to model steric and orientation effects, while the electronic activation energy is modeled using quantum mechanical (QM) simulations to calculate the energies of substrates and reaction intermediates for each potential SOM. This approach offers several advantages: • QM methods are based on fundamental physical principles and therefore transfer well between chemical classes; they do not rely on specific examples being present within a training set used to fit an empirical model • Each potential SOM is considered in the context of the whole molecular environment in which it resides, rather than identifying fragments within a substrate and treating

■

THEORY AND IMPLEMENTATION Modeling Principles. The key factors that determine the SOM are reactivity and accessibility. The models described herein estimate the activation energy at each potential SOM in a substrate using fundamental and transferable QM methods, rather than relying on empirical pattern matching with a limited domain of applicability. The QM models are independent of isoform, reflecting the consistent reaction pathways across isoforms. However, the binding pockets of the different CYP isoforms differ in size, shape, and chemical composition and influence the orientation of the substrate relative to the reactive oxy-heme core. Electrostatic, hydrogen-bonding and lipophilic interactions between substrate and CYP binding pocket have varying effects across isoforms and will cause different orientations to be favorable. In addition, the steric bulk within a substrate will influence the accessibility of sites to the reactive oxy-heme core, with those sites embedded toward the center being less accessible than those in open sites on the periphery of the substrate. These B

DOI: 10.1021/acs.jcim.6b00233 J. Chem. Inf. Model. XXXX, XXX, XXX−XXX

Article

Journal of Chemical Information and Modeling steric effects are also isoform specific, as the different sizes of the CYP binding pockets can accommodate different levels of steric hindrance. The models described herein are able to capture these orientation and steric effects with ligand-based models trained on isoform-specific data sets, enabling adjustments to be made to the QM-generated electronic activation energy that are specific to each isoform. The relative rate of metabolism for a site can be calculated from the activation energy, Ea, for the rate-limiting step in the reaction pathway, as the rate is proportional to the negative exponential of the activation energy (the Arrhenius equation): ki ∝ e−Eai / kT

Direct calculation of activation energies is computationally very expensive due to the need to perform a transition-state search. Instead, the heat of reaction, ΔHR, is calculated from the heat of formation of the substrate and reaction intermediates, and a Brønsted relationship24 is used to calculate an approximation to ΔHA (as a linear relationship has been shown to exist between the activation energy, ΔHA, and the heat of reaction, ΔHR). The parameters of the Brønsted relationship can be derived from detailed experimental regioselectivity data where available.16,25 However, in some cases there are insufficient experimental data and, instead, high level ab initio calculations can be used to accurately calculate the activation energies which, in turn, can be used to derive the parameters of the Brønsted relationship for the faster, semiempirical AM1 calculations.26,27 However, due to the differences in the chemical mechanisms and methods for calculation of each of the pathways leading to oxidation, the energy scales of the calculated activation energies will differ. Therefore, in order to compare the rates of reactions that proceed by different pathways, the activation energies must be on the same scale, and therefore a normalization must be applied. To achieve this, the activation energy scale relating to the abstraction of hydrogen from aliphatic carbon sites is used as a reference, and calculations performed with other methods are transformed onto this energy scale. Figure 2 shows the linear

(2)

where ki is the relative rate of metabolism, Eai is the activation energy of site i, k is the Boltzmann constant, and T is the temperature. As discussed above, by directly calculating the activation energy, the lability of sites can be compared on an absolute scale between compounds, rather than just a relative ranking of sites within a compound. This is achieved by comparing the rate of product formation with that of a decoupling pathway, to give an absolute assessment of the efficiency of the product formation step in the catalytic cycle. This enables a medicinal chemist to identify likely metabolically vulnerable positions in a molecule and can be used to guide development away from compounds with potentially rapid clearance. In the remainder of this section the various aspects of the CYP metabolism prediction models will be described in detail. Quantum Mechanical Models of Electronic Effects. The oxidizing species and the chemical mechanisms of oxidation are the same for all CYP isoforms. This allows the intrinsic vulnerability of the sites on a potential substrate to be calculated with reference only to the structure of that molecule. The oxidation reactions proceed via different pathways depending on the nature of the SOM on the substrate: aliphatic hydroxylation proceeds via an initial hydrogen abstraction followed by reaction of the ferryl oxygen with the alkyl radical, a process known as the rebound mechanism;17 alkene epoxidation proceeds via activation of the double bond to form an iron alkoxy radical species in a tetrahedral orientation;18 aromatic C-oxidation proceeds via activation of an aromatic bond19 followed by an intramolecular hydrogen atom transfer known as “NIH-shift”;20 and direct oxidation of heteroatoms such as sulfur and nitrogen proceeds via bond formation with the ferryl oxygen,21,22 although often dealkylation reactions are favorable over direct oxidation. The reactivity model performs a QM calculation to estimate the “electronic” activation energy, ΔHA, for every potential SOM, using knowledge of the reaction pathway for that site. ΔHA is the energy of the transition state relative to that of the reactants for the rate-limiting step of the reaction pathway. These calculations are performed using AM1,23 a QM approach based on a semiempirical Hamiltonian. [Note: Initial low-energy substrate geometries are generated with Corina40 as input to AM1 and further optimized within the AM1 calculations. Although the reaction energies can vary for different conformations of flexible substrates, the overall impact on the accuracy of prediction was found to be small. Therefore, for computational efficiency we consider only a single conformation for each substrate.] While less accurate than a full ab initio simulation, AM1 is many times faster. Ab initio simulations have been used to identify systematic errors due to the use of AM1, and correction factors are applied within the electronic model.15,16

Figure 2. Graph showing the linear relationship between H-abstraction activation energies calculated using DFT and the models based on AM1. The points represented by blue diamonds show the average ΔHA, estimated using a Brønsted relationship based on ΔHR calculated with AM1, plotted against the single activation energy value assigned to the corresponding sites in SMARTCyp, derived from DFT transition state calculations.15 While only a single activation energy is assigned to each class of site by SMARTCyp, in practice there may be significant variations between similar sites due to different molecular environments in which they occur. To illustrate this, the error bars show one standard deviation in the ΔHA values calculated using AM1 on the full molecules. These averages were calculated over a total of 2252 sites on a wide diversity of compounds. The minimum number of sites in each class for which an average is shown was 18. The transformation of N-oxidation and N-hydroxylation energies from DFT, on the basis of this linear relationship, is represented by the red squares.

relationship between the activation energies calculated using the ab initio B3LYP DFT method, as published by Rydberg et al.,15 and those estimated from ΔHR, calculated with AM1 and a Brønsted relationship, for hydrogen abstraction sites, as described below. Further details of the calculations performed for each of the reaction pathways modeled are provided in the next few subsections. C

DOI: 10.1021/acs.jcim.6b00233 J. Chem. Inf. Model. XXXX, XXX, XXX−XXX

Article

Journal of Chemical Information and Modeling Hydrogen Abstraction Reactions. The rate-limiting step in the formation of a metabolite by hydrogen abstraction has been identified as the abstraction of the hydrogen from the substrate by the oxy-heme and formation of an alkyl radical intermediate. In the Brønsted relationship used to estimate the activation energy, an additional linear term involving the ionization potential has also been found to be important to capture resonance effects in the transition state.25 Using detailed experimental measurements of the relative rates of product formation at different sites of the same molecule, this pathway was modeled as described in Korzekwa et al.25 to estimate ΔHA for each potential site of hydrogen abstraction. Aromatic Oxidation. Aromatic oxidation progresses by formation of a tetrahedral intermediate between the substrate and oxy-heme at the SOM, followed by rearrangement to form a hydroxylated product. The formation of the tetrahedral intermediate is the rate-limiting step in this process, and the activation energy was also found to be proportional to the heat of reaction. Using experimental measurements of the relative rates of formation of different metabolites on the same molecule, the parameters of this relationship can be determined and ΔHA calculated on the same scale as that for hydrogen abstraction, as described in detail by Jones et al.16 Double Bond Epoxidation. Similar to aromatic oxidation, the epoxidation of a carbon−carbon double bond proceeds via the formation of a tetrahedral intermediate, followed by rearrangement to form the epoxide.28 The formation of the tetrahedral intermediate is again the rate-limiting step, with the activation energy found to be proportional to the heat of reaction. There are insufficient experimental data with which to confidently parametrize a Brønsted relationship and, in this case, we rely on activation energies calculated with ab initio DFT calculations that were shown to agree with experimental observations, as described in Kumar et al.28 In this case, AM1 calculations of ΔHR exhibit an excellent correlation with the ab initio activation energies, as shown in Figure 3. This enables the

lowest value is used to estimate the relative rate of epoxidation of the corresponding double bond. Other Direct Oxidation Pathways. For S-oxidation, Noxidation, N-hydroxylation, and other pathways including desulfurization of phosphothioates, oxidation of disulfides, and aldehyde oxidation/deformylation, there are limited experimental data regarding the rates of these reactions relative to other sites on the same compounds. Furthermore, ab initio DFT calculations indicate that there is less variation in these rates between similar functionalities. Therefore, for these sites, activation energies derived from ab initio DFT calculations published by Rydberg et al.15 were transformed onto the same energy scale as the other sites described previously, using the linear relationship shown in Figure 2. As an illustration, the transformation of N-oxidation and N-hydroxylation energies using this linear relationship is also shown in Figure 2. Accessibility: Steric and Orientation Effects. In addition to the intrinsic vulnerability of a site in a molecule to oxidative attack, the accessibility of that site to the oxy-heme core will also influence the relative rate of metabolism. This effect is calculated as a correction to the activation energy due to the orientation of the molecule within the active site and steric hindrance by nearby atoms in the substrate. Orientation effects are modeled by descriptors representing the topological distance to important functionalities such as acidic, basic, hydrogen bond donor/acceptor, and lipophilic groups that interact with key residues in the CYP active site. The steric accessibility of a potential SOM depends on the surrounding atoms in the substrate and will be influenced by nearby bulky functionalities or whether the site is part of a ring, a conjugated system, or an aliphatic chain. The steric effects are modeled using descriptors representing the distance to functionalities introducing steric bulk surrounding the SOM. These functionalities are defined as SMARTS patterns,29 and some examples are given in Table 1. Table 1. Example Definitions of Functionalities from Which Distances Are Calculated as Steric and Orientation Descriptors Used To Model Accessibility of Potential Sites of Metabolism in Ligand-Based Models descriptor type

SMARTS

basic group

[N;+0;X3]

orientation orientation orientation

acidic group H-bond donor H-bond acceptor heavy atom ring atom

[\#8;H]-C=O [\#7,\#8,\#9;!H0] [N;!X4]

steric steric

Figure 3. Graph showing the linear relationship between ΔHR calculated with AM1 and ΔHA calculated with DFT for potential sites of double bond epoxidation, as published by Kumar et al.28

functionality

orientation

[!#1] [!#1,R]

description non-conjugated amine carboxylic acid intrinsic donor nitrogen acceptor non-hydrogen non-hydrogen in a ring

Isoform-specific data sets have been carefully curated from the literature, as described in more detail in the following subsection, with the steric and orientation descriptors calculated for all sites in all molecules. Principal component regression models30 were trained on these data sets using knowledge about the metabolic fate of each site to set the dependent variable: 0 for a non SOM, 1 for a primary SOM, 0.5 for a secondary SOM, and 0.25 for a tertiary SOM. In order to generate an adjustment to the electronic activation energy, ΔHA, it is necessary to calculate the adjustment on the same energy scale as ΔHA. This is achieved by including the ΔHA in the regression and scaling the descriptor

estimation of the DFT activation energy from the AM1 ΔHR, which, in turn, can be transformed to calculate ΔHA on the same energy scale as that for hydrogen abstraction, using the linear relationship shown in Figure 2. Epoxidation proceeds by formation of a tetrahedral intermediate with the carbon at either end of the double bond. Therefore, ΔHA is calculated for both potential sites, and the D

DOI: 10.1021/acs.jcim.6b00233 J. Chem. Inf. Model. XXXX, XXX, XXX−XXX

Article

Journal of Chemical Information and Modeling regression coefficients relative to the ΔHA regression coefficient. These scaled coefficients can then be applied to descriptors for new molecules and the resulting correction added to the “electronic” activation energy, ΔHA, to calculate the estimate of activation energy, Ea, adjusted for steric and orientation effects. Data Curation. A detailed review of the primary literature was performed to prepare high quality data sets of isoformspecific human CYP substrates annotated with SOM. The papers were manually parsed to extract primary, secondary, and tertiary SOM, along with the identity of the major and minor metabolizing CYP isoforms. The emphasis was on high-quality data, retaining only human data and excluding data generated with inappropriate experimental conditions, such as unphysiological substrate concentrations. The consequence of this is that the data sets are smaller than some of those previously published.9,31 However, analysis of the chemical space covered by the CYP data sets against launched drug space shows that good coverage of drug-like chemical space has been achieved, as illustrated in Figure 4, and the higher quality data are expected to

are available in the Supporting Information for inspection, including references to the primary literature from which the SOMs were identified. There is an element of judgement to be applied when classifying sites within a molecule as primary, secondary, or tertiary and identifying major/minor isoforms, with reliance placed on kinetic data from expressed supersomes and isoformspecific inhibition experiments with human liver microsomes. Variability between assays makes direct comparison of experimental data between publications challenging, but efforts have been made to make classifications across different molecules consistent. This was achieved by taking the consensus of multiple reviewers of the data, in addition to the views of the authors of the original paper, and looking for confirmation of the classifications across multiple assay protocols. Calculating Regioselectivity. The regioselectivity of metabolism is the proportion of metabolism that occurs at each site. This proportion is given by the rate of metabolism at that site relative to the sum of the rates for all potential SOM. The advantage of calculating an approximation to the activation energy is that a relative rate can be generated via eq 2, allowing the regioselectivity of metabolism at site i to be given by Ri =

result in more accurate models. Table 2 summarizes the number of compounds in the training set for each isoform, used to fit the contributions of the steric and orientation descriptors, and the independent test sets, used to validate each model. The data sets

Li =

a

Ntraining

Ntest

144 80 145 136 147 76 220

57 27 49 49 56 30 84

× 100

(3)

ki k water + ki

(4)

where kwater is the rate of water formation via the decoupling pathway. The value of kwater has previously been determined experimentally using an intrinsic isotope effect method34,35 for metabolism by CYP3A4, and therefore the lability values presented herein should be interpreted with reference to this isoform. The distribution of the lability of the sites on a compound is conveniently shown on a “metabolic landscape” histogram using color as a visual guide, based on the efficiency with which metabolism would occur at that site: red for labile (>0.80); yellow for moderately labile (between 0.35 and 0.80); green for moderately stable (between 0.05 and 0.35); and blue for stable (